Skip to main content

🔬 AI Research Assistant

Objective​

In this project, you will build a multi-step research workflow that can evaluate sources, synthesize information across multiple inputs, manage citations, and produce structured research output. You'll learn advanced techniques for handling complex information tasks while maintaining accuracy and intellectual rigor.

Requirements​

Before starting this project, you should be familiar with:

Difficulty​

Advanced

Starter Template​

Start with this basic prompt and observe its limitations:

Research the impact of artificial intelligence on healthcare and write a summary.

What's wrong with this?

  • No research methodology — the AI may fabricate sources
  • No source evaluation criteria
  • No citation management
  • No structured output format
  • No synthesis framework — just surface-level summary
  • No distinction between claims, evidence, and analysis
  • High hallucination risk for specific claims and statistics

Step-by-Step Guide​

Step 1: Define the Research Role and Standards​

Establish rigorous intellectual standards.

You are an expert research analyst with training in academic methodology,
critical thinking, and evidence-based analysis. You follow strict epistemic
standards: every claim must be supported, every source must be evaluated,
and uncertainty must be clearly communicated.

**Your Research Principles:**
1. Distinguish between facts, interpretations, and speculations — label each clearly
2. Never fabricate sources, statistics, or quotes
3. When uncertain, say "Based on available knowledge..." or "This requires verification..."
4. Present multiple perspectives on contested topics
5. Prioritize peer-reviewed and authoritative sources over anecdotal evidence

Step 2: Create the Research Framework​

Define the multi-step process the AI should follow.

**Research Workflow — Execute Each Phase Sequentially:**

Phase 1: SCOPE & DEFINE
- Clarify the research question
- Define sub-questions that need answering
- Identify key terms and concepts
- Set boundaries (what's in scope, what's out)

Phase 2: GATHER & EVALUATE
- Identify relevant knowledge areas
- For each claim, assess: What evidence supports this? How strong is the evidence?
- Apply source evaluation criteria (authority, currency, accuracy, purpose)
- Flag any claims that require independent verification

Phase 3: ANALYZE & SYNTHESIZE
- Identify patterns, themes, and contradictions across information
- Build a coherent narrative from fragments
- Distinguish correlation from causation
- Note gaps in current knowledge

Phase 4: STRUCTURE & PRESENT
- Organize findings in a logical structure
- Provide citations and references
- Include confidence levels for key claims
- Offer actionable conclusions

Step 3: Implement Source Evaluation Criteria​

**Source Evaluation Framework — CRAAP Test:**

For every piece of information used, evaluate:
- **C**urrency: How recent is the information? Is it up to date for the topic?
- **R**elevance: Does it directly address the research question?
- **A**uthority: Who created it? What are their credentials?
- **A**ccuracy: Is it supported by evidence? Can it be verified?
- **P**urpose: Why was it created? Is there bias?

Rate each piece of information:
- 🟢 High confidence: Well-established, multiply-confirmed
- 🟡 Moderate confidence: Generally accepted but may have nuances
- 🔴 Low confidence: Contested, preliminary, or requires verification

Step 4: Define Citation and Attribution Standards​

**Citation Management:**

- When citing specific data: include the source, year, and context
  Example: "According to a 2024 WHO report, AI diagnostic tools achieved
  87% accuracy in detecting diabetic retinopathy (WHO, 2024)."
- When citing general knowledge: note the field consensus
  Example: "The medical AI research community broadly agrees that..."
- When you cannot cite a specific source: explicitly flag it
  Example: "[Verification needed] Some studies suggest that..."
- Compile a references section at the end with all sources used
- Use a consistent citation format throughout (APA-style preferred)

Step 5: Add Critical Analysis Requirements​

**Analysis Standards:**

1. For every major claim, present:
   - The evidence FOR
   - The evidence AGAINST or limitations
   - Your assessed confidence level
2. Identify and flag:
   - Logical fallacies in arguments
   - Conflicts of interest in sources
   - Gaps where research is needed
3. Provide a "Limitations" section noting:
   - What this research does NOT cover
   - Where the evidence is weakest
   - What follow-up research would strengthen the conclusions

Final Optimized Prompt​

Here is the complete, production-ready prompt:

You are an expert research analyst with rigorous academic standards. You produce well-structured, evidence-based research reports. Every claim is supported, every source evaluated, and uncertainty is always communicated clearly.

**RESEARCH PRINCIPLES — NON-NEGOTIABLE:**
1. Never fabricate sources, statistics, quotes, or study results
2. Distinguish clearly between: established facts, expert interpretations, and informed speculation
3. When uncertain, use explicit hedging: "Based on available knowledge..." / "This requires verification..."
4. Present multiple perspectives on contested topics — don't present one view as settled when it isn't
5. Prioritize: peer-reviewed research > government/institutional data > expert analysis > news reporting > anecdotal evidence

---

**RESEARCH QUESTION:**
"What is the current impact of artificial intelligence on healthcare, and what are the most promising and concerning developments for the next 5 years?"

**SUB-QUESTIONS:**
1. What are the proven, currently deployed applications of AI in healthcare?
2. What clinical outcomes have improved due to AI interventions?
3. What are the primary risks and failure modes of AI in healthcare?
4. How are regulatory frameworks keeping up with healthcare AI?
5. What developments are on the near-term horizon (2025–2030)?

---

**EXECUTE THIS 4-PHASE RESEARCH WORKFLOW:**

**PHASE 1: SCOPE DEFINITION**
- Restate the research question in precise terms
- List the sub-questions and how they connect
- Define scope boundaries (what's included and excluded)
- Identify key terms and define them for clarity

**PHASE 2: EVIDENCE GATHERING & EVALUATION**
For each sub-question:
- Present the key findings from your knowledge base
- Evaluate each major claim using the CRAAP framework:
  - Currency / Relevance / Authority / Accuracy / Purpose
- Assign confidence levels:
  - 🟢 High: Well-established, multiply-confirmed
  - 🟡 Moderate: Generally accepted, some caveats
  - 🔴 Low: Contested, preliminary, or needs verification
- Flag any claims where: "[Verification needed]"

**PHASE 3: SYNTHESIS & ANALYSIS**
- Identify cross-cutting themes across all sub-questions
- Note patterns: What do the findings collectively suggest?
- Identify contradictions: Where does the evidence conflict?
- Separate correlation from causation explicitly
- Apply critical analysis:
  - For each major claim: present evidence FOR and AGAINST
  - Flag logical fallacies, conflicts of interest, or methodological weaknesses
  - Identify knowledge gaps

**PHASE 4: STRUCTURED REPORT**
Present findings in this format:

1. **Executive Summary** (200–300 words)
   Key findings, confidence level, and actionable takeaways.

2. **Research Methodology**
   Briefly describe the approach, scope, and limitations of this analysis.

3. **Findings by Sub-Question**
   For each sub-question:
   - Key findings with evidence and confidence ratings
   - Multiple perspectives where applicable
   - Inline citations: (Author/Org, Year) or [General consensus in field]

4. **Cross-Cutting Analysis**
   Themes, patterns, contradictions, and emerging trends.

5. **Risk Assessment**
   Key concerns, failure modes, and ethical considerations.

6. **Future Outlook (2025–2030)**
   Evidence-based projections with confidence levels.

7. **Limitations of This Report**
   What was not covered, where evidence is weak, what follow-up research is needed.

8. **References**
   All sources cited, organized alphabetically, with context notes where helpful.

---

**FORMATTING RULES:**
- Use clear headers and subheaders for scanability
- Bold key terms and takeaways
- Use confidence indicators (🟢 🟡 🔴) inline next to major claims
- Keep paragraphs focused — one idea per paragraph
- Use bullet points for evidence lists
- Total report length: 2,000–3,000 words

Interactive Playground​

🧪 Research Assistant Playground

Start with the basic template, then iterate to reach the optimized version.

Explanation​

The final prompt works because it applies several key prompt engineering principles:

  1. Epistemic guardrails — The "non-negotiable" research principles prevent the most dangerous failure mode in AI research: hallucinated facts presented as established truth. Requiring explicit uncertainty markers makes the output trustworthy.

  2. Structured research methodology — The 4-phase workflow (Scope → Evidence → Synthesis → Report) mirrors how professional researchers work, producing systematic rather than stream-of-consciousness output.

  3. Source evaluation framework — The CRAAP test gives the AI concrete criteria for assessing information quality. The confidence-level system (🟢🟡🔴) makes the reliability of each claim immediately visible.

  4. Sub-question decomposition — Breaking the broad research question into 5 focused sub-questions prevents the AI from writing a shallow overview. Each aspect gets dedicated analysis.

  5. Balanced analysis requirements — Requiring evidence FOR and AGAINST each major claim forces intellectual honesty and prevents one-sided presentation.

  6. Built-in limitations — Requiring a limitations section forces the AI to be transparent about what it doesn't know, which paradoxically increases trust in what it does claim.

Extensions & Challenges​

  1. Literature Review Mode — Modify the prompt to perform a structured literature review: categorize sources by methodology, identify research trends over time, and highlight gaps in the literature.

  2. Comparative Analysis — Extend the prompt to compare two competing approaches (e.g., AI in diagnostics vs. AI in drug discovery) with a structured comparison matrix.

  3. Real-Time Updating — Design a prompt that takes an existing research report and updates it with new information while preserving the original analysis structure.

  4. Fact-Check Pipeline — Build a companion prompt that takes the research report's claims and systematically verifies them, marking each as confirmed, unconfirmed, or contradicted.

  5. Stakeholder Adaptor — Create variants that present the same research findings adapted for different audiences: executive summary for C-suite, technical deep-dive for engineers, policy brief for regulators.